← Back to feed

A $37 GLM 5.3 red team: the Alloy-modeled auth layer held, but two bugs outside

AI · Security · 2026-09-09 10:26:00

After releasing identity management and role-based access control in GoodMem, the team red-teamed it with GLM 5.3. The layer is written in Java with a formal Alloy model describing how ownership, roles, grants and API-key restrictions combine. Using OpenCode's built-in agent with GLM 5.3 in a grey-box test, the agent could read source, schema and models but all attacks ran through the live public API. Two non-authorization bugs emerged: ordinary users could list all API-key metadata, and the USER role wrongly held extension-upload capabilities while the loader ignored disabled status, letting an attacker's JAR run with server privileges during a victim's retrieval. Memory instructions injected via a shared space also reached the victim's prompt unlabeled. Fixes shipped.

Source: goodmem.ai


Auto-generated summary for reference only. RSS